Accelerate Lunchtime Seminar Series
Starts: 2025/02/10 at 12:00
Ends: 2025/03/10 at 13:00
Join us to find out more about research taking place in AI for Science across the Accelerate Science community.
Details of future talks are available on Talks@Cam
Lunch provided, please register to attend via this form so we can confirm catering arrangements.
The ATLAS Virtual Research Assistant Heloise Stevance, Eric and Wendy Schmidt A.I. in Science Fellow, Department of Physics, University of Oxford
The ATLAS sky survey is able to image the whole night sky every 24 to 48 hours, looking for near earth asteroids and exploding stars. This generates 10s of millions of potential alerts every day which must be triaged and filtered to find the few explosions worth following up with more (expensive and time consuming) resources. Much of the work can be automated, but human “eyeballing” remains the final step before reporting and follow-up. This is a task that involves crappy images, sparse and uneven time series (with error bars and non-detections), and a whole lot of contextual knowledge to make sense of the mess. We are looking for rare events (not many training samples), we want (near) 100% recall, and explainability is paramount (since faults in the VRA can impact astrophysical event rates). That is a lot to ask of any model. In this presentation I will briefly present how ML is used in the VRA to lower eyeballer workload (decreased by 70%) and why “fancier” ML methods reported in the literature did not address our problems. I will also touch on upcoming sky survey challenges.
Distilling ML Models into Formulae for Ricci-Flat Metrics Viktor Mirjanic, PhD student, Department of Computer Science and Technology, University of Cambridge
Machine learning has shown great success in approximating Ricci-flat metrics on Calabi–Yau manifolds, but its black-box nature often limits interpretability. In this talk, I will show that for highly symmetric manifolds, the machine learning models used to approximate these metrics can be distilled into closed-form symbolic expressions. These expressions are compact, interpretable, and have the same accuracy as the original model.
These seminars are open to members of the University of Cambridge. For further details, please email accelerate-science@cst.cam.ac.uk.